Contrastive Representation Learning for Conversational Question Answering over Knowledge Graphs_2

2025-04-27 0 0 1.34MB 10 页 10玖币

侵权投诉

Contrastive Representation Learning for Conversational

estion Answering over Knowledge Graphs

Endri Kacupaj

kacupaj@cs.uni-bonn.de

University of Bonn, Germany

Kuldeep Singh

kuldeep.singh1@cerence.com

Zerotha Research and Cerence GmbH, Germany

Maria Maleshkova

maria.maleshkova@uni-siegen.de

University of Siegen, Germany

Jens Lehmann∗

lehmann@infai.org

Amazon

InfAI (Institute for Applied Informatics), Germany

ABSTRACT

This paper addresses the task of conversational question answering

(ConvQA) over knowledge graphs (KGs). The majority of exist-

ing ConvQA methods rely on full supervision signals with a strict

assumption of the availability of gold logical forms of queries to

extract answers from the KG. However, creating such a gold logical

form is not viable for each potential question in a real-world sce-

nario. Hence, in the case of missing gold logical forms, the existing

information retrieval-based approaches use weak supervision via

heuristics or reinforcement learning, formulating ConvQA as a

KG path ranking problem. Despite missing gold logical forms, an

abundance of conversational contexts, such as entire dialog history

with uent responses and domain information, can be incorpo-

rated to eectively reach the correct KG path. This work proposes

a contrastive representation learning-based approach to rank KG

paths eectively. Our approach solves two key challenges. Firstly,

it allows weak supervision-based learning that omits the neces-

sity of gold annotations. Second, it incorporates the conversational

context (entire dialog history and domain information) to jointly

learn its homogeneous representation with KG paths to improve

contrastive representations for eective path ranking. We eval-

uate our approach on standard datasets for ConvQA, on which

it signicantly outperforms existing baselines on all domains and

overall. Specically, in some cases, the Mean Reciprocal Rank (MRR)

and Hit@5 ranking metrics improve by absolute

and

points,

respectively, compared to the state-of-the-art performance.

CCS CONCEPTS

•Information systems →Question answering.

KEYWORDS

contrastive learning, conversations, question answering, KG

∗work done prior to joining Amazon

Permission to make digital or hard copies of all or part of this work for personal or

classroom use is granted without fee provided that copies are not made or distributed

for prot or commercial advantage and that copies bear this notice and the full citation

on the rst page. Copyrights for components of this work owned by others than ACM

must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,

to post on servers or to redistribute to lists, requires prior specic permission and/or a

fee. Request permissions from permissions@acm.org.

CIKM ’22, October 17–21, 2022, Atlanta, GA, USA

ACM ISBN 978-1-4503-9236-5/22/10. . . $15.00

https://doi.org/10.1145/3511808.3557267

Figure 1: Motivating example illustrating a sample conversa-

tion [10]. For conversational question answering over KGs,

the availability of the entire dialog history with uent re-

sponses and domain information acts as context source in

determining the ranking of KG paths while retrieving cor-

rect answers. Our proposed approach models conversational

context and KG paths in a shared space by jointly learning

the embeddings for homogeneous representation.

ACM Reference Format:

Endri Kacupaj, Kuldeep Singh, Maria Maleshkova, and Jens Lehmann. 2022.

Contrastive Representation Learning for Conversational Question Answer-

ing over Knowledge Graphs. In Proceedings of the 31st ACM International

Conference on Information and Knowledge Management (CIKM ’22), Octo-

ber 17–21, 2022, Atlanta, GA, USA. ACM, New York, NY, USA, 10 pages.

https://doi.org/10.1145/3511808.3557267

1 INTRODUCTION

Question answering over knowledge graphs (KGQA) is an essential

task that maps a user’s utterance to a formal query in order to

retrieve the correct answer [

]. Recently, with the increasing pop-

ularity of intelligent personal assistants (e.g., Alexa, Cortana), the

research focus of the scientic community has shifted to Conver-

sational Question Answering over KGs (ConvQA) with multi-turn

dialogues [10, 17, 33].

For the KGQA setup, the existing scientic-literature can be

broadly classied into two categories [

]: (i) semantic parsing

approaches, where the goal is to map questions into a logical form,

which is then executed over the knowledge graph to extract the cor-

rect answers. (ii) Information retrieval approaches aim to retrieve

a question-specic graph and apply ranking algorithms to select

arXiv:2210.04373v1 [cs.CL] 9 Oct 2022

CIKM ’22, October 17–21, 2022, Atlanta, GA, USA Endri Kacupaj, Kuldeep Singh, Maria Maleshkova, & Jens Lehmann

entities for top positions (i.e., rank KG paths). The two approaches

follow either a parse-then execute paradigm or a retrieval-and-

rank paradigm. For ConvQA over KGs, there has been signicant

progress on semantic parsing-based approaches [

]. How-

ever, collecting training data for semantic parsing approaches is

challenging, and time-consuming [

] since each question must

be associated with a gold logical form. While for the information-

retrieval/ranking-based approaches, only the correct answers (e.g.,

entities/KG paths) are required for each question.

State of the Art Limitations.

Existing ranking-based ConvQA

techniques formulate the task as a path ranking problem and pro-

pose either a heuristic approach [

] or a reinforcement learning

model [

] to learn from question reformulations (by solely rely-

ing on the user). However, these techniques have their inherent

limitations. For instance, the rule-based approaches generally re-

quire tedious manual labor to generate rules, are error-prone, and

are prone to bias [

]. Furthermore, existing scientic literature

points to several limitations in solely relying on users for query

reformulation, and focused on automatic query suggestion/related

query recommendations techniques [6, 7]. Firstly, entirely relying

on users creates a heavy cognitive load on them [

]. Secondly,

reformulation query issued within a time interval that is short

enough (such as ve minutes) implies dissatisfaction [

]. Lastly, a

recent study points out that critical issues for conversational sys-

tems concerning reformulations have not been widely analyzed

in the literature, especially the reformulations’ implicit ambiguity

[

]. Hence, for advancing IR-based ConvQA methods, there is a

desirable ask for approaches that are independent of availability

of gold-logical form, heuristic rules, or its sole dependency on the

users for query reformulation.

1.1 Proposed Approach, Motivation, and

Contributions

This paper addresses a relatively unexplored research area: the

ranking-based ConvQA task for answering conversational ques-

tions posed against a knowledge graph. In this context, we propose

PRALINE (

ath

anking for convers

tiona

quest

on a

ring),

a novel contrastive representation learning approach to rank KG

paths for retrieving the correct answers eectively. Furthermore,

we enrich the learning process by incorporating the available con-

versational context, i.e., (1) the entire dialog history with (2) uent

responses and (3) domain information (c.f. Figure 1). Our rationale

for critical choices are following:

Why Contrastive Learning for ConvQA?

Contrastive learning [

] aims at learning representations of

data by contrasting similar and dissimilar samples. For our task, the

data contains conversations and respective answers (label) for each

conversational question. However, there is no discrete information

on how the answers were extracted. Therefore fully supervised

learning approaches such as semantic parsing [

] cannot be

directly applied without annotations. Consequently, we can design

an approach with contrastive learning by only augmenting the data

with KG paths that lead to the correct answers. These KG paths have

as starting points the context entities mentioned in the conversation

and landing points the answers. Furthermore, extracted KG paths

leading to correct answers are marked as positive, while others

are negative. In this way, contrastive learning is ideal for our task

since it allows us to rank KG paths that are considered positive and

answer conversational questions.

Why Conversational Context to Enrich Learning Process?

Conversational context plays a vital role in human understand-

ing [

] and question answering [

]. To further enhance learning,

we seek additional conversational context to improve the ranking

performance and allow contrastive learning approach to distin-

guish between positive and negative paths. Such information can

be the conversation domain and uent natural language answers

[

] instead of standalone KG answers (answer labels).

Identifying the domain of the conversation allows us to enrich the

representations and eciently contrast negative paths that do not

use properties of the particular domain. Moreover, uent natural

language answers will supplement the conversations with addi-

tional textual context to support the learning and ranking process.

Contributions

: We make the following key contributions in the

paper: 1) We propose PRALINE, the rst contrastive learning based

approach for ConvQA that jointly models the available conversa-

tional context (full dialog history with uent responses and domain)

and KG paths in a common space for learning joint embedding repre-

sentations to improve KG path ranking. 2) We systematically study

the impact of incorporating additional context on the performance

of PRALINE. Results on standard datasets show a considerable im-

provement over previous baselines. To facilitate reproducibility and

reuse, our framework implementation and the results are publicly

available1. The structure of the paper is as follows: Section 2 sum-

marizes the related work. Section 3 provides the concepts, notations

and tasks denitions. Section 4 presents the proposed PRALINE

framework. Section 5 describes the experiments, including the ex-

perimental setup, the results, the ablation study and error analysis.

We conclude in Section 6.

2 RELATED WORK

Considering KGQA is a widely studied research topic, we stick to

the work closely related to our proposed approach (detailed surveys

are in [12, 43]).

Single-shot KGQA.

Several KGQA works handle the task as a se-

mantic graph generation and re-ranking. Bast and Haussmann

[1]

compare a set of manually dened query templates against the natu-

ral language question and generate a set of query graph candidates

by enriching the templates with potential relations. Yih et al

. [41]

creates grounded query graph candidates using a staged heuristic

search algorithm and employs a neural ranking model to score and

nd the optimal semantic graph. Yu et al

. [42]

use a hierarchical rep-

resentation of KG relations in a neural query graph ranking model.

Authors compare the results against a local sub-sequence alignment

model with cross attention [

]. Maheshwari et al

. [28]

conduct an

empirical investigation of neural query graph ranking approaches

by experimenting with six dierent ranking models. The proposed

approach is a self-attention-based slot matching model that exploits

the inherent structure of query graphs.

ConvQA over KGs.

Most recent works on ConvQA [

]

employ the semantic parsing approach to answer conversational

questions. The rst work in this area Saha et al

. [36]

propose a

1https://github.com/endrikacupaj/PRALINE

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ContrastiveRepresentationLearningforConversationalQuestionAnsweringoverKnowledgeGraphsEndriKacupajkacupaj@cs.uni-bonn.deUniversityofBonn,GermanyKuldeepSinghkuldeep.singh1@cerence.comZerothaResearchandCerenceGmbH,GermanyMariaMaleshkovamaria.maleshkova@uni-siegen.deUniversityofSiegen,GermanyJensLehman...

展开>> 收起<<

Contrastive Representation Learning for Conversational Question Answering over Knowledge Graphs_2.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Contrastive Representation Learning for Conversational Question Answering over Knowledge Graphs_2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: